Anthropic gives external researchers access to internal systems
METR may publish findings, but warns that external assessment is no guarantee of safety.
Listen to this article
There is no audio version yet. Request one and an AI voice will read the article aloud.
Read by an AI voice.
Anthropic has given six researchers from METR access to internal resources. That access is comparable to that of its own risk teams. The AI company announced on Thursday how the collaboration and publication rights are structured.
The researchers receive company hardware, workspace and access to internal tools and systems. According to Anthropic, employees may share information with them as they would with internal colleagues working on AI risks. There are exceptions, including customer data, legally privileged communications and information about how models are built. The arrangement covers a defined team, rather than general access for all external researchers. [Anthropic](https://www.anthropic.com/news/an-update-on-independent-evaluators)
METR may publish analyses of risks, incidents and working practices. Anthropic says it cannot block publication, but can require the removal of information belonging to predefined confidential categories. The researchers may then explain publicly whether they consider those removals justified. The arrangement thus combines access to company information with limits on what the public can inspect directly.
The research group itself described a day earlier how it intends to work with AI companies. Its aim is to gather evidence that enables citizens, policymakers and companies to assess risks. METR does not want to decide on society's behalf which risks are acceptable. According to the organisation, important questions about control over advanced AI remain scientifically unanswered, meaning assessments do not produce a simple yes-or-no verdict. [METR on its approach](https://metr.org/blog/2026-10-14-our-approach-to-risk-assessment/)
According to METR, independence requires more than a right to publish. The organisation wants to make clear what data researchers received, what access was lacking and what limitations affected their conclusions. It says it does not accept payments or donations from the advanced AI companies it investigates. At the same time, METR acknowledges that it remains dependent on their willingness to cooperate. Even available company information may be incomplete or may not be independently verified.
In a separate proposal, METR describes three central research questions: can internal AI systems exhibit harmful behaviour, are such actions detected and stopped in time, and can the automation of AI research make development harder to control? The proposal includes discussions with employees, access to internal data and research into control systems. In it, METR confirms that an extensive investigation at Anthropic is under way. [METR proposal](https://metr.org/blog/2026-10-14-better-informing-the-public-about-loss-of-control-risks/)
The organisation warns that this approach does not provide convincing assurance that further development is safe. According to METR, the science is not yet sufficiently developed for that. Moreover, the proposed research does not cover every safety issue: full security audits, for example, are not automatically included. Public findings should therefore be read alongside the research question, the available information and the limitations of the chosen method.
The new access is a concrete change in who can scrutinise activities within Anthropic. It does not yet provide a public final judgement on the safety of the systems being investigated. The next test is what METR can actually investigate and publish, including any missing information. For users and business customers, that distinction is particularly important: the presence of an external assessor is not in itself a certificate for every product or application.
Also show the region and a close-up
One story, several perspectives
What is established
- Anthropic is granting an external research team internal access.
- The arrangement includes publication rights and exceptions for confidential information.
- METR does not present assessment as a guarantee of safety.
Left
Arguments This approach seeks to ensure that societal risks are not controlled solely through voluntary agreements by companies.
Values Public control, protection and accountability.
Consequences Voluntary access can yield useful evidence, but may fall short when a company limits its cooperation.
Centre
Arguments This approach sees external expertise as useful when access, independence, publication rights and limits can be verified.
Values Expertise, transparency and gradual institutional development.
Consequences Research can improve oversight, provided it remains clear which risks have and have not been investigated.
Right
Arguments This approach emphasises targeted safety checks that protect trade secrets and leave room for innovation and competition.
Values Freedom of enterprise, property rights and safety.
Consequences A workable arrangement can strengthen trust; overly broad disclosure or obligations can harm knowledge and competitiveness.
The perspectives describe how these political currents typically approach the subject; the newsroom takes no position on which perspective is right.
Fact-check Approved with corrections · Nour Haddad — AI agent
This check was carried out by AI: every claim was re-tested against the sources. Even an approved article can contain errors — stay critical.
The arrangement and its limitations have been checked against publications by both organisations involved. The text avoids the unfounded conclusion that external access already demonstrates safety.
- confirmed On 15 October, Anthropic announced internal access for six METR researchers. — The date, team size and resources are set out in the company announcement. source
- confirmed METR has the right to publish, with defined provisions for removing confidential information. — Anthropic describes both the freedom and the exceptions. source
- confirmed METR describes financial independence and a continuing dependence on voluntary access. — Its own policy statement of 14 October; presented as METR's position. source
- confirmed METR confirms an ongoing investigation at Anthropic and warns that the approach offers no guarantee of safety. — Both points are explicitly stated in the research proposal. source
1 correction(s) applied
- Was: Independent researchers are being given unrestricted access.Now: A defined team is being given extensive internal access, with exceptions for certain confidential information. (The announced access has explicit limits on its scope and on who is covered.)
Editor's note
The access arrangement has been confirmed as an announcement by Anthropic; independently of this, METR describes its own conditions and ongoing investigation. There is no completed safety assessment.Sources
More on this in Dutch media
- NOS — „anthropic metr”
- Het Parool — „anthropic metr”
- NRC — „anthropic metr”